Like I said earlier, I think this Votecastr thing is crap, but if she's actually winning Iowa and Ohio, then there's good reason to think that Arizona, Georgia, and possibly Alaska are in play.
After all that, Wang and Silver ended up in the same place. Wang's median forecast is HRC 307, and Silver's probabilistic average (or median?) is HRC 302.
Trump has won. Dominated the rust belt, and Florida, while still holding all Romney states. Even came close in some states thought to be out of play for him, e.g. in VA despite a McMullin vote of around 60K. The only question is how the popular vote will play out. Soul-searching time for pollsters. Well, I was wildly off for predicting Clinton would win in an electoral landslide of 322-216. At least I thought her odds were ~85% and not 100%.
All those statisticians who said Clinton had a 99% chance of winning look like real ********ing morons today. All rise for King Bill Mitchell.
I gave it a 25% chance. Which still looks low and I still was shocked, but a bit closer than some of these models.
I apologize to @appoo and @MatthausSammer for calling them diaper babies. I should have been more of a diaper baby myself. I trusted the polls and the models and they showed a somehow clear victory for Hillary. I guess that Nate Silver was right to make his model sensible to intangibles and swings, making it the less optimistic model.
Apology accepted, but I'm not sure I saw this coming either. I think this election has major implications for the polling industry and the forecasting industry that's built around it. Most everyone got it wrong, and even Silver's model didn't necessarily pinpoint the Rust Belt as a concerning area for Clinton despite being correct regarding pumping the brakes on the certainty of a Clinton win. In the aftermath I look forward to much smarter people than myself diving into the data looking at why they were wrong and what indicators were flashing warning signs that they didn't see.
As Silver pointed out, the number of undecided voters made actually forecasting the horserace much more difficult.
A recession happening under a Republican president, Obama's personal magnetism and charisma, Clinton's vote for the Iraq War, and Palin swung Obama into office. Perhaps the ground game was overvallued in Obama's victory.
I still believe a ground game can matter when people are enthusiastic about your candidate and/or message.
Donald Trump won 76% of counties w/ a Cracker Barrel & 22% of counties w/ a Whole Foods -- a 54% gap. In '92, gap b/t same counties was 19%.— Dave Wasserman (@Redistrict) November 9, 2016 How's them numbers?
I worked with Silver for a few years before he went to BP full time and then on to fame via polling. I've never had a peek under the hood of the model, but I can guess based upon work experience what even he couldn't get a handle on. Here goes: What Nate did right was he understood correlation between states and the introduction of additional risk. He incorporated that in addition to wider confidence intervals around national polling. He also understood that national polling errors will have an outsized influence on certain states. They are more responsive to overall changes due to demographics in much the same way stocks have different betas. If the market goes up 5%, stock with beta = 2 would be expected to increase 10% while stock w/ beta of 0.5 would be expected to increase 2.5%. This is why his confidence probabilities were lower than other aggregators and what the nerd fight was about. His last article went to great pains to explain this and he also included a chart indicating what a generic 3 pt shift either way would do to state forecasts, recognizing why some states margins would shift greater than 3 pts based upon demographics. This was his Trump path to victory. As it turns out, the margins were about 3 pts off nationally...they may be less as we get the absentees compiled in CA, but we'll call it 3. Silver was projecting that certain states would swing maybe 1.1-1.2x of the national swing. Many swung more: WI 2.1x PA 1.6x MI 1.5x NC 1.5x MN 1.5x NH 1.1x FL 0.7x The issue in question is the source of both the national swing and how that translates to states. There could be all sorts of scenarios in which the polls are generally off by 2-3 pts and third party undecideds may not even be significant. Lack of response/dishonest response, which may have been a factor with women. A big one is not understanding/misestimation of the likely voter composition in terms of demographics. -It could be many small things, like AA turnout is a couple points lower, older up a couple pts, whites vote a couple pts more for X, Hispanics a couple more for Y. When a bunch of small things occur, states tend to move with the country. You can't have greater proportion of all demos. Everything sums to 1. So the volatility of states vs. country tends to be pretty clsoe to 1. -It could be one thing. Like whites generally move 5 pts. This has more volatility. -Worse case, it's a combination of only 2 or maybe 3 things. States that contain sizeable populations of the 2-3 key variables are going to be much more responsive to national polling errors. This is what happened. The biggest issue was older working age Whites really moving away from Dems. Typically, the 65+ crowd votes more GOP than 45-64. Compared to 65+ voter margins in 2008, 45-64 group moved 15 pts or so to the right in WI, MI, and PA. This was overall, not whites, but it reflects the working class white vote. People probably filled with anxiety because theyve seen 20 years of wage stagnation, retirement age coming up and no money. Another was AA turnout. So what we are looking at was (1) National polls wrong, (2) why they were wrong at the demo level and (3) how that translates to the model, which says, "There's a 3 pt swing here, it could be for any combination of these reasons, so we will weight the impact of each of these scenarios to estimate how much WI moves when the country moves 3 pts." A bad (but not worse case) nat swing of 3pts combined with a very bad case of what was driving that swing...both in terms of demos and the relatively small margins in PA/WI/MI created the problem. I don't think it was complete luck. You don't go looking for EVs where the margins are insanely high. The Rust Belt was the obvious place to look if demographics was creating bridges too far in VA/NM/CO/NV. If there is a non-response issue for women in that group, I don't know what pollsters can do about it. If they are fitting the responses here to a likely voter model under the assumption of turnout of x% and x+5% turn up, I don't know what they can do about that either. The models/polls were built to some extent to reflect traditional political alignments and those alignments are based upon demographics. voting patterns within demos, and some assumption of turnout for each group. They typically don't change that much every 4 years.